We propose a novel speech separation model designed to separate mixtures with an unknown number of speakers. The proposed model stacks 1) a dual-path processing block that can model spectro-temporal patterns, 2) a transformer decoder-based attractor (TDA) calculation module that can deal with an unknown number of speakers, and 3) triple-path processing blocks that can model inter-speaker relations. Given a fixed, small set of learned speaker queries and the mixture embedding produced by the dual-path blocks, TDA infers the relations of these queries and generates an attractor vector for each speaker. The estimated attractors are then combined with the mixture embedding by feature-wise linear modulation conditioning, creating a speaker dimension. The mixture embedding, conditioned with speaker information produced by TDA, is fed to the final triple-path blocks, which augment the dual-path blocks with an additional pathway dedicated to inter-speaker processing. The proposed approach outperforms the previous best reported in the literature, achieving 24.0 and 23.7 dB SI-SDR improvement (SI-SDRi) on WSJ0-2 and 3mix respectively, with a single model trained to separate 2- and 3-speaker mixtures. The proposed model also exhibits strong performance and generalizability at counting sources and separating mixtures with up to 5 speakers.
In this paper, we propose an efficient and accurate streaming speech recognition model based on the FastConformer architecture. We adapted the FastConformer architecture for streaming applications through: (1) constraining both the look-ahead and past contexts in the encoder, and (2) introducing an activation caching mechanism to enable the non-autoregressive encoder to operate autoregressively during inference. The proposed model is thoughtfully designed in a way to eliminate the accuracy disparity between the train and inference time which is common for many streaming models. Furthermore, our proposed encoder works with various decoder configurations including Connectionist Temporal Classification (CTC) and RNN-Transducer (RNNT) decoders. Additionally, we introduced a hybrid CTC/RNNT architecture which utilizes a shared encoder with both a CTC and RNNT decoder to boost the accuracy and save computation. We evaluate the proposed model on LibriSpeech dataset and a multi-domain large scale dataset and demonstrate that it can achieve better accuracy with lower latency and inference time compared to a conventional buffered streaming model baseline. We also showed that training a model with multiple latencies can achieve better accuracy than single latency models while it enables us to support multiple latencies with a single model. Our experiments also showed the hybrid architecture would not only speedup the convergence of the CTC decoder but also improves the accuracy of streaming models compared to single decoder models.
Reliable automatic hate speech (HS) detection systems must adapt to the in-flow of diverse new data to curtail hate speech. However, hate speech detection systems commonly lack generalizability in identifying hate speech dissimilar to data used in training, impeding their robustness in real-world deployments. In this work, we propose a hate speech generalization framework that leverages emotion knowledge in a multitask architecture to improve the generalizability of hate speech detection in a cross-domain setting. We investigate emotion corpora with varying emotion categorical scopes to determine the best corpus scope for supplying emotion knowledge to foster generalized hate speech detection. We further assess the relationship between using pretrained Transformers models adapted for hate speech and its effect on our emotion-enriched hate speech generalization model. We perform extensive experiments on six publicly available datasets sourced from different online domains and show that our emotion-enriched HS detection generalization method demonstrates consistent generalization improvement in cross-domain evaluation, increasing generalization performance up to 18.1% and average cross-domain performance up to 8.5%, according to the F1 measure.
Diffusion models are a new class of generative models that have recently been applied to speech enhancement successfully. Previous works have demonstrated their superior performance in mismatched conditions compared to state-of-the art discriminative models. However, this was investigated with a single database for training and another one for testing, which makes the results highly dependent on the particular databases. Moreover, recent developments from the image generation literature remain largely unexplored for speech enhancement. These include several design aspects of diffusion models, such as the noise schedule or the reverse sampler. In this work, we systematically assess the generalization performance of a diffusion-based speech enhancement model by using multiple speech, noise and binaural room impulse response (BRIR) databases to simulate mismatched acoustic conditions. We also experiment with a noise schedule and a sampler that have not been applied to speech enhancement before. We show that the proposed system substantially benefits from using multiple databases for training, and achieves superior performance compared to state-of-the-art discriminative models in both matched and mismatched conditions. We also show that a Heun-based sampler achieves superior performance at a smaller computational cost compared to a sampler commonly used for speech enhancement.
Social media platforms serve as accessible outlets for individuals to express their thoughts and experiences, resulting in an influx of user-generated data spanning all age groups. While these platforms enable free expression, they also present significant challenges, including the proliferation of hate speech and offensive content. Such objectionable language disrupts objective discourse and can lead to radicalization of debates, ultimately threatening democratic values. Consequently, organizations have taken steps to monitor and curb abusive behavior, necessitating automated methods for identifying suspicious posts. This paper contributes to Hate Speech and Offensive Content Identification in English and Indo-Aryan Languages (HASOC) 2023 shared tasks track. We, team Z-AGI Labs, conduct a comprehensive comparative analysis of hate speech classification across five distinct languages: Bengali, Assamese, Bodo, Sinhala, and Gujarati. Our study encompasses a wide range of pre-trained models, including Bert variants, XLM-R, and LSTM models, to assess their performance in identifying hate speech across these languages. Results reveal intriguing variations in model performance. Notably, Bert Base Multilingual Cased emerges as a strong performer across languages, achieving an F1 score of 0.67027 for Bengali and 0.70525 for Assamese. At the same time, it significantly outperforms other models with an impressive F1 score of 0.83009 for Bodo. In Sinhala, XLM-R stands out with an F1 score of 0.83493, whereas for Gujarati, a custom LSTM-based model outshined with an F1 score of 0.76601. This study offers valuable insights into the suitability of various pre-trained models for hate speech detection in multilingual settings. By considering the nuances of each, our research contributes to an informed model selection for building robust hate speech detection systems.
We present 3DiFACE, a novel method for personalized speech-driven 3D facial animation and editing. While existing methods deterministically predict facial animations from speech, they overlook the inherent one-to-many relationship between speech and facial expressions, i.e., there are multiple reasonable facial expression animations matching an audio input. It is especially important in content creation to be able to modify generated motion or to specify keyframes. To enable stochasticity as well as motion editing, we propose a lightweight audio-conditioned diffusion model for 3D facial motion. This diffusion model can be trained on a small 3D motion dataset, maintaining expressive lip motion output. In addition, it can be finetuned for specific subjects, requiring only a short video of the person. Through quantitative and qualitative evaluations, we show that our method outperforms existing state-of-the-art techniques and yields speech-driven animations with greater fidelity and diversity.
Confidence estimation, in which we estimate the reliability of each recognized token (e.g., word, sub-word, and character) in automatic speech recognition (ASR) hypotheses and detect incorrectly recognized tokens, is an important function for developing ASR applications. In this study, we perform confidence estimation for end-to-end (E2E) ASR hypotheses. Recent E2E ASR systems show high performance (e.g., around 5% token error rates) for various ASR tasks. In such situations, confidence estimation becomes difficult since we need to detect infrequent incorrect tokens from mostly correct token sequences. To tackle this imbalanced dataset problem, we employ a bidirectional long short-term memory (BLSTM)-based model as a strong binary-class (correct/incorrect) sequence labeler that is trained with a class balancing objective. We experimentally confirmed that, by utilizing several types of ASR decoding scores as its auxiliary features, the model steadily shows high confidence estimation performance under highly imbalanced settings. We also confirmed that the BLSTM-based model outperforms Transformer-based confidence estimation models, which greatly underestimate incorrect tokens.
Recently, multi-stage systems have stood out among deep learning-based speech enhancement methods. However, these systems are always high in complexity, requiring millions of parameters and powerful computational resources, which limits their application for real-time processing in low-power devices. Besides, the contribution of various influencing factors to the success of multi-stage systems remains unclear, which presents challenges to reduce the size of these systems. In this paper, we extensively investigate a lightweight two-stage network with only 560k total parameters. It consists of a Mel-scale magnitude masking model in the first stage and a complex spectrum mapping model in the second stage. We first provide a consolidated view of the roles of gain power factor, post-filter, and training labels for the Mel-scale masking model. Then, we explore several training schemes for the two-stage network and provide some insights into the superiority of the two-stage network. We show that the proposed two-stage network trained by an optimal scheme achieves a performance similar to a four times larger open source model DeepFilterNet2.
We propose a multi-task learning (MTL) model for jointly performing three tasks that are commonly solved in a text-to-speech (TTS) front-end: text normalization (TN), part-of-speech (POS) tagging, and homograph disambiguation (HD). Our framework utilizes a tree-like structure with a trunk that learns shared representations, followed by separate task-specific heads. We further incorporate a pre-trained language model to utilize its built-in lexical and contextual knowledge, and study how to best use its embeddings so as to most effectively benefit our multi-task model. Through task-wise ablations, we show that our full model trained on all three tasks achieves the strongest overall performance compared to models trained on individual or sub-combinations of tasks, confirming the advantages of our MTL framework. Finally, we introduce a new HD dataset containing a balanced number of sentences in diverse contexts for a variety of homographs and their pronunciations. We demonstrate that incorporating this dataset into training significantly improves HD performance over only using a commonly used, but imbalanced, pre-existing dataset.
In recent years, end-to-end speech recognition has emerged as a technology that integrates the acoustic, pronunciation dictionary, and language model components of the traditional Automatic Speech Recognition model. It is possible to achieve human-like recognition without the need to build a pronunciation dictionary in advance. However, due to the relative scarcity of training data on code-switching, the performance of ASR models tends to degrade drastically when encountering this phenomenon. Most past studies have simplified the learning complexity of the model by splitting the code-switching task into multiple tasks dealing with a single language and then learning the domain-specific knowledge of each language separately. Therefore, in this paper, we attempt to introduce language identification information into the middle layer of the ASR model's encoder. We aim to generate acoustic features that imply language distinctions in a more implicit way, reducing the model's confusion when dealing with language switching.